Web infrastructure for AI agents · v1.0


One API key to fetch any page as Markdown, search the live web, pull brand kits and DESIGN.md files, plus the utilities agents keep reaching for. Every answer is cached at the edge, and cache hits are free.

0
endpoints behind one key
0
credits for any cache hit
0
search providers, auto failover
MCP
native, same key, every tool
01How it works


Each call is canonicalized, cached in three tiers and returned in one typed envelope, so your agent spends tokens on the task, not on scraping.
FETCH01 / 04

Read

Any URL to clean Markdown. Native extraction first, a headless reader when a page needs JavaScript.

› fetch blog.cloudflare.com
# The Cloudflare Blog
## Latest posts
Durable Objects now ship with
SQLite storage built in...
stripped: nav, ads, 14 links
SEARCH02 / 04

Find

Live web results through six providers with automatic failover, so one outage never reaches your agent.

› query: "durable objects"
langsearchok
jinastandby
automstandby
exastandby
+ brave, duckduckgo
BRAND03 / 04

Recognise

A domain in, a brand kit out: ranked logos, a palette with roles, typography and company facts.

› brand stripe.com
primary#533afd
fontsohne-var
DESIGN04 / 04

Describe

Any site as a DESIGN.md: tokens, type scale and components your coding agent can build against.

---
name: "Linear"
theme: dark
radius: 8px
---
## Typography
Inter Display, 48/1.0, -0.02em
02Live console


The first call does the work. The second comes back from the edge in milliseconds, and costs nothing. Pick an endpoint and press run.
req_brand_00
POST/v1/brand { "domain": "stripe.com" }
Pipeline
  1. Resolve stripe.com, SSRF guard
  2. Render homepage, collect assets
  3. Rank 7 logos, dedupe by hash
  4. Extract palette with roles
  5. Wikidata company facts
  6. Attach DESIGN.md
Response
Press Run to send the request.
latency...
cache...
credits...
providernative + wikidata

Sample responses, trimmed. Latencies are typical production figures.

Run it for real in the playground
03Caching


Inputs are canonicalized (tracking params stripped, hosts lowercased, keys sorted), so equivalent requests share one entry. Bound staleness per call with max_age.
  • Brand, cold2,140 ms
  • Search, cold1,812 ms
  • Fetch, cold384 ms
  • Global KV hit31 ms
  • Edge hit6 ms

Typical production latencies on Cloudflare Workers · November 2026

0×
faster on an edge hit than a cold brand lookup
0
credits for any response served from cache
0 + SWR
cache tiers, with stale-while-revalidate on top
L0
In-flight coalescing
Identical concurrent requests share one upstream call.
L1
Edge cache
Per-location Cache API, read before anything else.
L2
Global KV
Expensive results replicate to every Cloudflare location.
SWR
Stale-while-revalidate
Stale entries answer instantly while a refresh runs behind them.
04Endpoints


Core endpoints do the heavy lifting. Utilities are small, reliable lookups agents call constantly, at one credit each.
EndpointBest forFresh forProvidersCredits
Fetch/v1/fetchAny URL to clean Markdown, HTML or text.1hEdge fetch, Jina Reader1
Search/v1/searchRanked live web results, optionally with page content.15mLangSearch, Jina Search +42
Design.md/v1/designA site's design system as a DESIGN.md file.7dEdge fetch3
Brand/v1/brandLogos, colors, fonts, socials and company facts for a domain.7dEdge fetch, Wikidata +15
Map/v1/mapDiscover every URL on a domain.1dEdge fetch2
Weather/v1/weatherCurrent conditions and daily forecast for any place.10mOpen-Meteo1
Exchange rates/v1/fxReference FX rates and conversion, latest or historical.1hFrankfurter (ECB)1
DNS/v1/dnsResolve DNS records over HTTPS.5mCloudflare 1.1.1.11
WHOIS/v1/whoisRegistration data for domains and IPs.1dRDAP1
Wikipedia/v1/wikiEncyclopedia summary for any topic.1dWikipedia1
Public holidays/v1/holidaysPublic holidays for 100+ countries.30dNager.Date1
Stock quote/v1/stockLive stock quote and company profile.1mFinnhub via Composio1
05Docs


  • RESTPlain JSON over HTTPS. One envelope, one error shape, every endpoint.
  • MCPOne server for Claude, Cursor or your own agent loop, same key.
  • UIAdd ?ui=true and get a json-render spec your app can draw directly.
  • LLMSOpenAPI and llms.txt so agents can read the docs themselves.
curl https://actionlayer.fiftyseven.is/v1/brand \
  -H "Authorization: Bearer $ACTIONLAYER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"domain":"stripe.com"}'
200 OK · x-cache: HIT · 0 creditsFull reference
06Security


01
Hashed at rest

Keys are shown once, stored only as a peppered HMAC-SHA256, and checksummed so malformed keys never reach the database.

02
Scoped

Per-key scopes, optional expiry, one-click roll and revoke.

03
Header-only

Keys in URLs are refused, so they never land in access logs.

04
SSRF-guarded

Private ranges, internal hosts and odd ports are blocked on every hop.

05
Metadata-only logs

Request inputs and outputs are never stored.

06
SOC 2 coming soon

A SOC 2 Type II audit is on the roadmap. No report has been issued yet; the security page lists what is in place today.

07Pricing


Credits are only spent when we do upstream work. Anything served from cache is on us, on every plan.
Free
$0/month
  • 1,000 credits every month
  • 60 requests per minute per key
  • Up to 5 active keys
  • Every endpoint, REST and MCP
  • Cached responses always free
ProMost teams
$29/month
  • 100,000 credits every month
  • 600 requests per minute per key
  • Up to 25 active keys
  • Every endpoint, REST and MCP
  • Cached responses always free
Credits per uncached call: Fetch 1 · Search 2 · Design.md 3 · Brand 5 · Map 2 · Weather 1 · Exchange rates 1 · DNS 1 · WHOIS 1 · Wikipedia 1 · Public holidays 1 · Stock quote 1
08Start building


Get your API key1,000 free credits every month. No card required.
09FAQ


What is ActionLayer?

An API that gives AI agents and apps the live web: page fetching as Markdown, search, brand kits, DESIGN.md extraction and a set of everyday utilities, behind one key and one MCP server.

Why are cache hits free?

A hit costs us almost nothing to serve, so it costs you nothing. Equivalent requests are canonicalized to share an entry, and you can bound staleness per call with max_age.

Does it work with Claude, Cursor and other MCP clients?

Yes. Point any MCP client at /mcp with your key in the Authorization header and every endpoint shows up as a tool.

What happens when a search provider goes down?

The next healthy provider answers. A failing provider is skipped for a minute, or fifteen on quota and auth errors, so your agent never waits on it.

Do you store what my agents fetch?

No. We keep request metadata for your logs and billing. Inputs and outputs live only in the cache, keyed by a hash, until they expire.

How much does ActionLayer cost?

The Free plan includes 1,000 credits a month with no card required. Pro is $29 a month for 100,000 credits. Cache hits cost 0 credits on every plan.

Who operates ActionLayer?

Valiant Digital Technologies LLP, based in Mumbai, India. Support is at support@actionlayer.fiftyseven.is.